Is Smaug Humanizer state of the art?Pending benchmarks.Announced Oct 26, 2026; no independent results yet.

The details so far

Lab

Abacus.AI

Announced

Oct 26, 2026

API

Not yet · None announced; self-hosted (vLLM, llama.cpp, Ollama, LM Studio)

Open weights

Published · Hugging Face: abacusai/Smaug-Humanizer (BF16) and abacusai/Smaug-Humanizer-GGUF (Q8_0 fits a 16 GB GPU; Q4_K_M runs on an 8 GB GPU)

Parameters

12B (11,959,730,224 safetensors parameters)

Type

Text rewriting model: rewrites AI-generated text into a human style

Base model

google/gemma-4-12B (the pretrained model, not instruction-tuned)

Training

SFT with LoRA (r=64, alpha=128, 2 epochs) on 3,676 AI-to-human rewrite pairs, then 200 GRPO steps with a kernel-similarity (MMD) style reward and five deterministic gates (numbers, layout, length, script, spelling)

License

Apache 2.0

Usage

Split long inputs into chunks of about 500 words; sampling temperature 1.0, top_p 0.95, top_k 64, stop on <end_of_turn>

Languages

Mostly English; other languages work less well, per the model card

Tags marked unsure are not stated on the lab's own pages.

What would let us call it

Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.

• Publicly announced with a source
• Usable by the public (API or weights)
• Listed on an independent board we track
• Enough results for a rank in a category
• Holds across two snapshots (crowns only)

What the lab claims

Self-reported. Shown for context; not used in any verdict.

Lab claim: turns many texts that the Pangram AI-text detector scores as AI into text it scores as human, though not every time; long texts are flagged more often.

Sources